Latest / In the Loop AI / Kimi K2.5: The Best New Model is Open-Source (Again!)

Transcript

Kimi K2.5: The Best New Model is Open-Source (Again!) · In the Loop AI

0:00
Josh: Over the last few weeks, one of the hottest new topics that exists in the world of AI has been agents.
0:05
Josh: Agents that can actually get into your computer and do things for you.
0:08
Josh: We've seen this with Claude Cowork, which can actually access files and make
0:11
Josh: changes on your computer.
0:12
Josh: And then to the fullest extent, we just recorded an episode on Claude Bot,
0:16
Josh: which allows an actual computer to fully take over your life,
0:20
Josh: send messages on your behalf, take care of emails, book reservations.
0:24
Josh: The problem with that is that even an afternoon of use can cost hundreds of dollars.
0:28
Josh: So what we've done today is we've actually figured out a
0:31
Josh: way to replace that totally for free where
0:34
Josh: you get the same quality outputs but with none of the cost and
0:37
Josh: to do that i'm going to start on anthropics website because the
0:41
Josh: reality is that the screen that you're seeing right now isn't actually anthropics
0:44
Josh: website in fact it was built using this new tool completely for free in about
0:47
Josh: 25 minutes which i thought was such an amazing demo it was built using a tool
0:51
Josh: called kimmy k 2.5 which is the newest model coming out of china that is fully
0:55
Josh: open source fully open weight and in order to build this all i had to do was
1:00
Josh: feed it a video so you'll see on the screen here i
1:03
Josh: generated a video on my desktop using a screen recorder that copied the Anthropic
1:08
Josh: website, I said, hey, just take the screen recording of the website and create
1:11
Josh: an exact replica of the website for me.
1:13
Josh: 25 minutes later, without any additional prompts, it listed all the things it did.
1:17
Josh: It went through all the design and then actually published a full preview of
1:20
Josh: the website that we can see here on this read record.
1:22
Josh: So this model is incredible. I don't know if you've had a chance to play around
1:25
Josh: with it or check it out, but this is a huge change in the world of agents because
1:29
Josh: of how capable it is for such a low cost.
1:32
Ejaaz: This model is trending pretty heavily online right now.
1:36
Ejaaz: I actually saw someone describe Moonshot Labs, the creator of this model,
1:41
Ejaaz: as the anthropic of China.
1:43
Ejaaz: This was a quiet release, Josh. So the creators of this model kind of just updated
1:49
Ejaaz: their chatbot interface with Kimi K 2.5 and didn't tell anyone.
1:55
Ejaaz: And within a few hours of that launch, remember, no publicist or anything,
1:58
Ejaaz: It was the number one trending model on Hugging Face, which is like where everyone
2:02
Ejaaz: goes to access all these free open source AI models. And...
2:07
Ejaaz: What you just demonstrated, I think, is one of the core reasons why this model is so special.
2:12
Ejaaz: So to give a few stats about this, it was trained on about 15 trillion tokens.
2:18
Ejaaz: And typically, AI models are trained on text tokens specifically.
2:22
Ejaaz: This wasn't the case with Kimike 2.5. It was trained on text and audio and visual
2:28
Ejaaz: and a bunch of other mediums.
2:30
Ejaaz: And the reason why this is important is it allows you to do the example that
2:33
Ejaaz: you just show, Josh, which is feed it a video or in this case,
2:37
Ejaaz: a screen recording of a website that you wanted to build and build it in exactly that way.
2:41
Ejaaz: And the reason why this is important is it shifts the use of AI models from
2:46
Ejaaz: explaining what you need to do to it.
2:48
Ejaaz: So like, hey, could you do this, like describing what you want from it into
2:53
Ejaaz: just showing it what you want to build.
2:55
Ejaaz: And I think that that's like a really intuitive way for people to interact with
2:59
Ejaaz: AI models versus like people that aren't just quite literate like me sometimes
3:03
Ejaaz: when I'm trying to explain something. Right.
3:04
Ejaaz: The second really cool thing is you started off this episode mentioning agents, Josh.
3:10
Ejaaz: And I think this is really important because KimiK 2.5 has this superpower where
3:15
Ejaaz: they can spin up up to 100 sub-agents.
3:19
Ejaaz: Think of a sub-agent as just another instance or replica of KimiK 2.5,
3:23
Ejaaz: but it's specifically focused on doing a certain task.
3:27
Ejaaz: So for example, if your goal is to figure out whether investing in Anthropic is a good idea.
3:33
Ejaaz: It'll spin up one agent that does the research, another agent that does the
3:36
Ejaaz: fact checking, another that tests different kind of architectures.
3:40
Ejaaz: And the cool part about this is they can work in parallel, which means that
3:44
Ejaaz: you can cut down the execution time for a task by four and a half times.
3:49
Ejaaz: So imagine you had a task that took four and a half hours, you can now do it in one hour.
3:53
Ejaaz: And I think this kind of like multi-agent trend that you identified or that
3:57
Ejaaz: you spoke about is super important because that's what we're seeing with the
4:01
Ejaaz: likes of Anthropic with Cloud Code and Cowork and OpenAI with Codex.
4:06
Ejaaz: But the fact that this thing is free is completely insane. Josh,
4:10
Ejaaz: do you know how much it cost, rumored, for them to train this?
4:13
Josh: I have no idea, but I would imagine a tremendous amount of money for them to
4:17
Josh: just train it and then release it fully open source, open weight.
4:21
Ejaaz: So the rumor, and again, this is not fact. I wish I could fact check this.
4:25
Josh: And also, to be fair, before you say this, the Chinese models are notorious
4:28
Josh: for lying about how much it costs. Yes, they are. So take this number with a grain of salt.
4:32
Ejaaz: You're right. So the number that's being floated is $4.6 million,
4:37
Ejaaz: which is nothing. That seems so low.
4:40
Ejaaz: It seems so low, which is nothing compared to the billions of dollars that OpenAI
4:45
Ejaaz: has spent to kind of train their models.
4:46
Ejaaz: And to give you guys an idea, like why we're comparing Kimi-K 2.5 to these like
4:51
Ejaaz: frontier AI models built by OpenAI and Anthropic is because in some cases,
4:55
Ejaaz: it's almost as good as this.
4:57
Ejaaz: Like if you look at its performance on humanity's last exam,
5:00
Ejaaz: which is notoriously the hardest benchmark for an AI model to be tested on,
5:03
Ejaaz: it scored a 50.2%, which beats Claude's latest model, Opus 4.5, and GPT 5.2.
5:12
Ejaaz: It doesn't quite beat Anthropic at coding, Josh.
5:16
Ejaaz: I know you built that cool website in a few minutes, and which makes me think
5:19
Ejaaz: that maybe it's really good at front end development but just a really impressive
5:23
Ejaaz: model and i'm guessing it's like super cheap to operate as well compared to
5:26
Ejaaz: like some of these expensive.
5:26
Josh: Models yeah we're going to get into the cost because if you do want to use
5:29
Josh: it at length and you don't have a couple h100 gpu sitting in your your home
5:33
Josh: you're going to have to pay a little bit uh thankfully it's significantly less
5:37
Josh: and we'll get into the prices but one thing you mentioned is that it's actually
5:39
Josh: not the best coding model in the world and i think that's okay that's not the
5:43
Josh: real breakthrough one of the most amazing breakthroughs is actually before we
5:47
Josh: were recording the show, Ijaz, you showed a demo of,
5:49
Josh: you gave CloudCode the website of Figma and said, hey, can you go emulate this
5:53
Josh: website? And it actually did a pretty good job.
5:55
Josh: The difference between CloudCode and something like this new model is that
5:59
Josh: I was able to feed it just a video. And what it did is it analyzed each frame,
6:03
Josh: each pixel within each frame, understood the context of each pixel,
6:06
Josh: and then figured out how to intuitively regenerate that in a webpage using code
6:11
Josh: and like whatever type of design tools that it used.
6:13
Josh: And that is the novel thing because most models do image to code,
6:17
Josh: but Kimi K 2.5 does video to understanding to code. And I think that's one of
6:22
Josh: the more novel breakthroughs.
6:23
Josh: One of the three, actually. The second with it being natively multimodal.
6:27
Josh: So you mentioned 15 trillion tokens that it was trained on, but it's mixed between
6:31
Josh: visuals and text for the first time.
6:33
Josh: So this really has a good understanding of videos, of photos.
6:36
Josh: It's starting to even get the Google physics.
6:37
Josh: And then the third part that she mentioned, which is the Asian Swarm.
6:40
Josh: I want to spend a little bit of time on this because the Asian Swarms are super cool.
6:43
Josh: We actually have an example of one of the Asian Swarms and how it works.
6:47
Josh: The way it works is it's able to separate...
6:49
Josh: Itself into basically a hundred small mini tasks and the example that
6:52
Josh: we're seeing on screen now is a film script and
6:55
Josh: it's a short story that the model generated it created
6:58
Josh: a shot list it created renderings of images of
7:01
Josh: what this the frames of this could look and it generated basically
7:04
Josh: an entire movie in a fraction of the time that it
7:07
Josh: would take to do as a single model i think the
7:10
Josh: actual number is four and a half times faster like
7:13
Josh: you mentioned earlier versus traditional models so it can call up to 1500 tools
7:18
Josh: it's like this swarm of agents working on a single problem so it's faster it's
7:22
Josh: more efficient it can it's just like you can make a movie script in five minutes
7:27
Josh: and it'll generate the entire thing for you with a shotless and all i mean some
7:31
Josh: of these examples are pretty unbelievable did
7:32
Ejaaz: You hear about the the underlying mechanism that they use to to build this it's
7:37
Ejaaz: actually super cool because well one thing chinese ai model labs are repeatedly known for doing is,
7:43
Ejaaz: you know, they don't get access to all these fun, expensive GPUs that the Western labs get to.
7:48
Ejaaz: So they have to get really creative in their research and training techniques.
7:52
Ejaaz: And they did that with the agents.
7:54
Ejaaz: So to get that four and a half times efficiency that you mentioned,
7:57
Ejaaz: they use this technique called parallel agent reinforcement learning.
8:02
Ejaaz: Typically, when you spin up 100 agents.
8:04
Ejaaz: You're going to have a hard time. And the reason why you're going to have a
8:07
Ejaaz: hard time is something called agent collapse.
8:09
Ejaaz: So typically, a model will be used to doing things in sequence.
8:14
Ejaaz: So if you ask it to do a really complex task, it's going to start with task
8:18
Ejaaz: one and only proceed to task two once it's done with task one.
8:22
Ejaaz: And if you spin up a bunch of agents, the model might sometimes still do things
8:27
Ejaaz: sequentially. And you don't want that. You want it to do in parallel.
8:30
Ejaaz: So this new training technique that they spun up and pioneered,
8:33
Ejaaz: there's a paper all about this, is super unique and never been done before,
8:37
Ejaaz: which allows them to not get model collapse at 100 agents.
8:40
Ejaaz: The crazier part is each of these agents get access to over 1,500 tools.
8:46
Ejaaz: And that's what makes an agent useful. You go from an LLM telling you what is
8:50
Ejaaz: useful to an agent that can actually do something on your behalf.
8:53
Ejaaz: That's pretty impressive. And then the final thing is they have this thing called,
8:57
Ejaaz: well, I actually don't know what it's called technically, but the way I understood
9:01
Ejaaz: it is it's kind of like a brain. It's called an orchestrator.
9:03
Ejaaz: And so it ingests the tasks that you've asked it to do, and it breaks it down
9:08
Ejaaz: into multiple different tasks.
9:09
Ejaaz: The fact that it could do it for this cheap, Josh, sorry, I know I keep mentioning
9:13
Ejaaz: the cost, but you need to tell the people the cost because it's just insane.
9:17
Ejaaz: This is something that I would use regularly, a company would use regularly
9:20
Ejaaz: because it saves them so much money.
9:21
Josh: And we actually have this really cool visual of the orchestrator on screen here,
9:25
Josh: which gives you a visual representation of what that looks like.
9:28
Josh: So the orchestrator breaks this down into sub-agents. It assigns them tasks.
9:31
Josh: And then the tasks kind of go back and forth through a fact checker.
9:35
Josh: There's a file downloader. There's a web developer. There's this entire toolkit.
9:38
Josh: So one of these is emulating an AI researcher. The other is a physics researcher.
9:42
Josh: The other is a life science researcher.
9:43
Josh: And what you're getting is a series of experts across every domain working on
9:48
Josh: problems in parallel with access to 1,500 of these tools, like the fact checker,
9:52
Josh: the file downloader, the web developer,
9:54
Josh: the text scraper, so it can view images and understand and interpret what they
9:58
Josh: mean. And it's such a powerful...
10:00
Josh: Stack that you have. And without the collapse, with this software novelty that
10:03
Josh: they've introduced, it allows them to do this unbelievable thing.
10:07
Josh: So to your point, when, I mean, historically, China has been hardware constrained
10:11
Josh: and they've really accelerated on the software.
10:13
Josh: And this is very much an extension of that acceleration. Now we have some more
10:16
Josh: examples that are very fun to show that I would love to show because as I was
10:20
Josh: going through to prepare for the episode this morning, I was like,
10:23
Josh: wow, this is pretty cool.
10:23
Josh: And EJ, you even dropped in one of your own that I thought was pretty neat.
10:27
Josh: So if you don't mind explaining what this one is here.
10:28
Ejaaz: Well, it may not be the coolest example, but this is something that I would personally do.
10:33
Ejaaz: And I know a bunch of my friends would do in their spare time,
10:35
Ejaaz: which is just again, on websites, the fidelity of these things is pretty insane.
10:41
Ejaaz: And I have to emphasize, we're going from a screen recording to like a fully
10:47
Ejaaz: functional front end development.
10:49
Ejaaz: And I don't think people quite understand how necessarily hard it is to do front end development.
10:54
Ejaaz: I think a lot of software engineers that do this will kind of scoff at that comment.
10:57
Josh: But it is true.
10:58
Ejaaz: It is like super hard to do because there's the design element,
11:01
Ejaaz: which is incredibly subjective, which is Kimi K2.5's exact point.
11:05
Ejaaz: Instead of trying to describe it to an LLM, you can just kind of take a screen
11:08
Ejaaz: recording and spin this up in a matter of minutes, right? It took you, I think, 7.5 minutes.
11:15
Ejaaz: I just want to like emphasize a point that you made earlier before this,
11:19
Ejaaz: Josh, which is the agent side of this model is super important because if you
11:25
Ejaaz: look at a model from Anthropic,
11:27
Ejaaz: their flagship product is Claude Code and recently Claude Cowork.
11:32
Ejaaz: If I were to tie both of those products in one unique trait,
11:37
Ejaaz: it's the fact that you can spin up multiple agents.
11:40
Ejaaz: In fact, the founder of Claude Code, and in fact, a lot of the Anthropic team
11:44
Ejaaz: do between 80 to 100% of production level code.
11:48
Ejaaz: So that means new products that they ship completely built by Claude Code.
11:52
Ejaaz: Now, they're not doing this using one model.
11:54
Ejaaz: They're doing this spinning up several instances of.
11:58
Ejaaz: Kind of like put this into perspective this is
12:01
Ejaaz: the future of software development and software development pretty
12:05
Ejaaz: much underpins any major major breakthrough for
12:08
Ejaaz: any industry going forwards software and tech underpins everything
12:12
Ejaaz: so if you have an ai model that costs a fraction of the amount that the frontier
12:17
Ejaaz: flagship model from anthropic does and is 100 free and open source yeah you
12:22
Ejaaz: might need whatever 50k to 100k's worth of gpus to run it on your own instance
12:26
Ejaaz: but you can get access to Kimi K2.5's API right now.
12:29
Ejaaz: That is a huge advantage. And the Chinese AI labs just somehow stay on top.
12:34
Ejaaz: I don't know how they do this.
12:36
Josh: Yeah, well, if we're talking about token price, maybe we could get into the economics first.
12:39
Josh: We'll skip ahead a little bit because I think the economics of this is super
12:43
Josh: interesting, where if you don't have access to the GPUs in your house,
12:46
Josh: which I'm guessing nobody listening to this episode really does,
12:49
Josh: if you don't work at a major AI lab, well, there are ways in which you can run
12:52
Josh: this for free or close to free.
12:54
Josh: Now, the example that i showed earlier where i created a website clone
12:57
Josh: that was free because kimmy gives you three agentic
13:01
Josh: tasks per week basically that you can use for free but after you've exhausted
13:04
Josh: that there are some economic pricing sheets that we can use to kind of compare
13:09
Josh: this to other models in terms of cost so for opus 4.5 which is the most popular
13:13
Josh: model that we've been using everyone's been using it's fairly expensive the price input
13:18
Josh: Per token per million tokens is five dollars
13:21
Josh: while the output is 25 dollars so for
13:24
Josh: every million tokens you generate with opus 4.5 which is
13:26
Josh: the flagship coding model it costs 25 bucks for kimmy k 2.5 the input is 60
13:32
Josh: cents per million tokens and the output is three dollars that is almost a full
13:37
Josh: order of magnitude 90 decrease in price relative to opus 4.5 at a very comparable rate.
13:45
Josh: And that's just if you're comparing it on code. If you're comparing it on general
13:48
Josh: agentic tasks, it's actually slightly more capable than Opus 4.5 for one-tenth of the cost.
13:53
Josh: So if you're using something like CloudBot, which we recorded an amazing episode
13:56
Josh: on earlier this week, which you should go check it out, you can just swap in
13:59
Josh: this new model and run it through whatever cloud service you want.
14:03
Josh: And the price of your agent will be one-tenth of the cost.
14:07
Josh: And this happened in the matter of a couple of days. So the costs are rapidly decreasing.
14:11
Josh: And I think that advantage of it one being open source but two
14:13
Josh: being cost effective is huge for everyone if you remember
14:16
Josh: i think it was two or three weeks ago anthropic cut off
14:19
Josh: xai from using cloud code they actually removed
14:23
Josh: access to it and because it's closed
14:26
Josh: source there's nothing they could do about it but if they're using a model like kimmy
14:28
Josh: k2 to run k2.5 to run
14:31
Josh: their agents to build their code there's no one who can actually sever that
14:35
Josh: tie and it's the same for developers where if you're building on a platform
14:38
Josh: and you don't want it to change well now you have the open weights it's going
14:41
Josh: to be locked in forever it's going to be a fraction of the cost this is a really
14:44
Josh: viable substitute for those who are loving the agentic life os workflows
14:50
Ejaaz: Do you have any idea how they're able to produce this for such cheap costs?
14:56
Ejaaz: Because I'm trying to rack my brain around this, right? So, okay,
14:59
Ejaaz: sure, you've made a few research breakthroughs.
15:02
Ejaaz: The Chinese labs in particular are known for discovering or commercializing
15:07
Ejaaz: mixture of experts, which kind of cut down prompting and inference costs and
15:12
Ejaaz: training in general to a fraction of the price.
15:14
Ejaaz: But still, they don't have access to some of the top hardware, right?
15:18
Ejaaz: And kind of like Moore's Law would state that eventually a bunch of these GPUs
15:23
Ejaaz: that are A-grade are going to
15:25
Ejaaz: cost so much less and run like 100x more inference performance per token.
15:31
Ejaaz: So it's going to cost a lot less in general over time.
15:34
Ejaaz: They don't have access to these resources that the West does,
15:36
Ejaaz: right? So you mentioned like Anthropics cost, right?
15:38
Ejaaz: You said what? It was like $5 in and how much out?
15:41
Josh: $25 per million tokens compared to three.
15:44
Ejaaz: Okay, that used to be $15 in and $75 out when they launched the product.
15:49
Ejaaz: So we've come down by a significant factor since then. But again,
15:53
Ejaaz: I would imagine they did this because of cheaper, more effective chips.
15:57
Ejaaz: How has Kimi K2.5 done that? How has Moonshot done that?
16:00
Josh: I think I have two answers to this question. The first is through software innovation.
16:04
Josh: I assume they have cracked some sort of a code that allows them to generate less tokens per output.
16:09
Josh: The second one is the margins. Ejaz, how much money did Anthropic make last year?
16:13
Josh: It was what, $10 billion of revenue?
16:15
Ejaaz: $10 billion, correct.
16:17
Josh: China is unfortunately not the leader in AI and therefore they need every incentive
16:21
Josh: in the world to dethrone the leader of AI.
16:24
Josh: One of the ways you could do that is by winning on the margins and cutting down
16:28
Josh: those margins for your competitors.
16:30
Josh: Clearly they have this strategy because they're publishing this open source,
16:33
Josh: open weight. And you're seeing that happen with the pricing as well.
16:36
Josh: I assume a large part of that revenue
16:39
Josh: from Anthropic is just margin on the inference that they're charging.
16:42
Josh: It doesn't cost them anywhere near $25 per million tokens, but they're able
16:46
Josh: to charge for it because they're the leading frontier model that all of these
16:49
Josh: labs and businesses are willing to pay in order to use their services.
16:52
Josh: In the case of Kimi K2.5, they don't care. They don't need to make profit.
16:56
Josh: They just want them in market share.
16:58
Josh: And to do that, they're able to undercut pretty aggressively here. Like I'm sure...
17:02
Josh: Anthropic could match this and perhaps not actually lose money.
17:05
Josh: But that profit thing is real.
17:07
Ejaaz: It also helps that they have an absolute gigabrain as their founder and CEO.
17:11
Ejaaz: I don't know if you've looked into this guy, but this dude is only 31 years old.
17:16
Ejaaz: He was born in China. He went to Tsinghua University, which is actually the
17:22
Ejaaz: most popular university for AI and ML researchers in the world to graduate from.
17:27
Ejaaz: 50% of the world's top AI researchers, by the way, reside in China,
17:30
Ejaaz: and a large chunk of them graduated from Tsinghua.
17:33
Ejaaz: But Josh, he also did his PhD at Carnegie Mellon and he did it in under four
17:38
Ejaaz: years in assumedly robotics and machine learning, which is very impressive.
17:42
Ejaaz: And he also did a very long stint building out Google Brain and meta AI research.
17:48
Ejaaz: So he was probably one of those meta researchers getting paid tens of millions
17:51
Ejaaz: of dollars a year. So this guy's track record is insane.
17:54
Ejaaz: So it doesn't, I guess, with that CV, doesn't kind of surprise me that he's
17:58
Ejaaz: made these breakthroughs somehow. even on the hardware that he's constrained.
18:02
Josh: Yeah, it's incredibly impressive. I'd love to hear more from them.
18:05
Josh: In fact, we actually, the first time we heard from the founder was earlier today
18:08
Josh: with the announcement post. I had never really seen what he looked like.
18:11
Josh: I hadn't really heard him communicate.
18:12
Josh: It feels like it's a very sheltered, kind of quiet, secretive workplace that they have there.
18:18
Josh: But I'm hopeful that we'll start to see more because my God,
18:20
Josh: the talent there must be unbelievably impressive.
18:23
Josh: Just in China in general, when we talk a lot about the trading competitions
18:26
Josh: that we have, China's always seemingly winning.
18:28
Josh: They're just, they're doing really well and clearly they have incredible talent
18:31
Josh: density now you're showing on screen something that i'm very excited to talk
18:35
Josh: about which is jumping back to the examples of what you can actually do with
18:38
Josh: this new model and one of them is this really fun blueprint to 3d model space
18:43
Josh: now ijez you've watched friends right this might look familiar to you oh
18:46
Ejaaz: Yeah uh not not the uh the one on the left yeah not the exact uh high.
18:51
Josh: Fidelity you don't know the blueprint of the
18:55
Josh: Yeah, it's pretty cool. So it took a two-dimensional blueprint of a room and
19:00
Josh: it generated a three-dimensional version of Monica's apartment or Monica and Rachel's apartment.
19:05
Josh: I haven't watched Friends, but I know it's very popular and I've seen clips
19:08
Josh: from this room. So I'm familiar which one it is.
19:09
Josh: And it's a testament to the types of new creative things that you can do now
19:13
Josh: that it has the image to critical thinking to output.
19:17
Josh: Uh type of thinking process through generating these outputs and i just thought
19:21
Josh: that was really interesting there's a lot of really fun use cases that you can use and
19:25
Ejaaz: Dude this is a this is a ten thousand dollar a month apartment at minimum josh
19:29
Ejaaz: it's it's making me feel poor looking at the schematic oh my god yeah right right.
19:33
Josh: Growth street new york city that might even be more than 10 grand that's prime real estate come
19:37
Ejaaz: On dude yeah that's insane how how were they able to pay rent they were making
19:41
Ejaaz: comedic jokes the entire time for for seven years.
19:44
Josh: But also this becomes a very useful tool for real
19:47
Josh: estate agents right because they want to kind of recreate spaces
19:50
Josh: allow you to feel and live in the space more and
19:53
Josh: granted this is a low fidelity version but i'm sure this is step one in
19:56
Josh: creating some higher fidelity mock-ups of spaces that you would possibly want
20:00
Josh: to rent if you're building a house if you're building anything this is great
20:03
Josh: for construction for modeling these services used to cost a ton of money for
20:06
Josh: virtual renderings now they're effectively free or very close to it maybe just
20:10
Josh: a couple cents per output and that decrease is pretty substantial open
20:15
Ejaaz: Source is having quite the.
20:16
Josh: Week they're having a moment i've
20:17
Ejaaz: Commented a lot about this before but um i've said
20:20
Ejaaz: that i i never think open source will actually ever catch
20:23
Ejaaz: up to frontier level uh capabilities and in this case in some ways it does in
20:29
Ejaaz: some ways it doesn't um josh you know uh in my period or era of life right now
20:33
Ejaaz: i am a coding agent maxi i'm incredibly bullish on anthropic so uh you know
20:38
Ejaaz: i i scrutinize any other competitor pretty heavily when it comes down to this.
20:42
Ejaaz: I don't think it is as good as CloudCode. You mentioned this earlier,
20:46
Ejaaz: but it's scarily good in some aspects, right? With the front end development.
20:50
Ejaaz: So I'm curious to see how people use this. And I think what I love most about
20:54
Ejaaz: this is a lot of my friends that kind of want to do more creative pursuits,
20:58
Ejaaz: like build websites and do more front end stuff.
21:00
Ejaaz: They don't want to pay 200 bucks a month, CloudCode max, right?
21:04
Ejaaz: But they can get this for free and they can access it today.
21:07
Ejaaz: You literally built your website today, like in a few minutes before the The
21:11
Ejaaz: show starts and then recorded it.
21:12
Josh: In 25 minutes with one prompt.
21:13
Ejaaz: That's that's insane that's insane so if you can do it if i can do it anyone
21:17
Ejaaz: else listening to this can do it definitely go give it a go like i want to see
21:20
Ejaaz: some examples that people kind of like do with this.
21:23
Josh: The kimmy website itself actually has a bunch of
21:25
Josh: ideas and use cases that you can use to kind of
21:28
Josh: emulate or get inspired by and this is one of the major things
21:31
Josh: with the model launch like the reason we're talking about this today is
21:34
Josh: because they provide this really awesome demo online of them
21:37
Josh: screen recording a website and then emulating that and
21:40
Josh: creating it in five minutes and that's what we did and that's why it's so
21:43
Josh: exciting so the models that are not only able to make it accessible
21:46
Josh: through lower cost pricing but to kind of give you these curated
21:49
Josh: experiences where you can satisfy some sort of goal that you want in a way that's
21:53
Josh: easy all i asked was hey just create a clone of this website do it identically
21:57
Josh: and don't make any mistakes and it did it in one shot i think that is a critical
22:01
Josh: threshold required to onboard a lot more people to be excited to use this stuff
22:05
Josh: to go and set up quad bot it's pretty technically challenging
22:09
Josh: It takes a little while. It's not for the faint of heart. But something like
22:11
Josh: this, where they give you these use cases, they make it available for basically
22:14
Josh: free Mium, where you can pay extra if you want to use it more.
22:18
Josh: It's really exciting to see. And yeah, I mean, China's totally having a moment.
22:22
Josh: And open source is totally having a moment. Like both of these things are converging
22:26
Josh: at once to create all of the news this week, while the major AI labs who are
22:29
Josh: closed source are just kind of working in silence, perhaps trying to figure
22:33
Josh: out how to best react to something like this that becomes open source and available.
22:37
Josh: Now, you have to imagine, EJS, it's been a little while since we got a new big
22:40
Josh: dog on the block, a new Frontier model for one of these labs.
22:43
Josh: So the silence is deafening, but generally, the longer the silence goes,
22:48
Josh: the bigger the boom that follows.
22:50
Josh: And I suspect we are only a few weeks away from some new models that will make
22:54
Josh: this Kimmy K2.5 look like child's play, which is crazy to see.
22:58
Josh: Because right now it feels unbelievable and magical, but I'm sure it is soon
23:02
Josh: to be dethroned when the new models come out. So I guess we'll be here to follow
23:06
Josh: along with all of that news.
23:07
Josh: Ejos, any final thoughts before we part today?
23:10
Ejaaz: I mean, once again, the clear winner for all of this is the users.
23:14
Ejaaz: Big time. We get access to all these frontier models for either a cheap or free
23:18
Ejaaz: option. It's super cool.
23:20
Ejaaz: Or if you want to pay the extra amount and get like a curated experience,
23:23
Ejaaz: you can also do that as well.
23:24
Ejaaz: If you want to use a Chinese AI model, go for it. If you want to use a Western lab, pick your poison.
23:31
Ejaaz: What I'll finish up with is the pace of development for these things, Josh, is
23:38
Ejaaz: So underrated. Like, I feel like we are so spoiled. When we first started this
23:43
Ejaaz: show around like eight months ago.
23:45
Josh: We were like, oh man,
23:46
Ejaaz: Like it can produce a pretty good market summary of this investment,
23:51
Ejaaz: but like it's nothing like crazy.
23:53
Ejaaz: Fast forward to today and I'm reading a tweet on my timeline from the founder
23:57
Ejaaz: of Claude saying like, yeah, 100% of the code that we make, aka every new product
24:01
Ejaaz: that we build going forward is managed by Claude, like is managed by Anthropic.
24:06
Ejaaz: And I can assume that with a product like KimiK 2.5, they're probably doing the same thing.
24:12
Ejaaz: So are we entering the era where AI just builds itself? Probably.
24:17
Ejaaz: Super scary. I read an essay last night, word to the wise, don't read scary
24:22
Ejaaz: essays at night, where Dario Amode, founder of Anthropic, wrote about his bearish
24:26
Ejaaz: thesis and why we need to be super careful going forwards because we're entering AGI, dare I say.
24:32
Ejaaz: I don't know. I'm super excited. These model developments are super cool.
24:35
Ejaaz: And I'm excited for Josh Codex is probably going to come up with an upgrade
24:41
Ejaaz: OpenAI's new coding model is coming out in the next couple of weeks I'm excited
24:44
Ejaaz: they're having a town hall today fingers crossed that they probably want to announce it but maybe
24:48
Ejaaz: but when they do this will be the first platform to hear it on,
24:52
Ejaaz: now I know a bunch of you have listened to this and are thinking hmm,
24:55
Ejaaz: I'm going to download Kimmy K2.5 or just use it and test it out.
24:59
Ejaaz: I have a task for you to try out.
25:01
Ejaaz: In fact, it involves not one, but two sub-agents.
25:05
Ejaaz: Number one, ask it what the top AI show is on YouTube or any favorite platform
25:10
Ejaaz: that you listen to or hear on.
25:13
Ejaaz: And then ask it to subscribe if you aren't. Turn on notifications and give it a five-star rating.
25:18
Ejaaz: I have asked this of you for the Claude Clawbot episode.
25:23
Ejaaz: I'm going to ask it for you for any of the Kimmy K 2.5 fans out,
25:27
Ejaaz: please support us. It helps us massively.
25:30
Josh: Yeah, if you ever need a use case, you can just have it. Go figure out how to
25:32
Josh: subscribe autonomously to the YouTube channel.
25:34
Ejaaz: That'd be pretty cool.
25:36
Josh: And share it with your 10 closest friends through iMessage once you get hooked
25:39
Josh: up with ClawBot. That would be great.
25:41
Josh: But yes, all of these cool, exciting new things that you were talking about,
25:44
Josh: including Dario's Anthropic Letter
25:46
Josh: and the OpenAI State of the Union that they're kind of hosting today.
25:50
Josh: We're going to cover that on our episode later this week in the AI Roundup. So stay tuned for that.
25:54
Josh: And yeah, we'll see you guys in that episode. Thanks for watching.